摘要：AbstractThe reliability of data centers can be severely affected when battery failure occurs in the Uninterruptible Power Supply (UPS). Thus it has become a central issue for the industry to discover failure-impending batteries in UPS. In this paper, we consider this important problem and present a data-driven method for predictive battery maintenance. The major contributions are as follows.First, we develop a changepoint detection technique for efficient data labeling. Second, new features are designed to fully utilize the dataset. Third, we build a predictive classification model which can discriminate between healthy and failure-impending batteries. Our method has been built and evaluated on 209,912,615 records from Tencent data center involving nearly 300 batteries monitored over 2 years. The experiment on test set shows that our method is able to predict battery replacement with 98% accuracy and averagely 15 days in advance, which outperforms the previous maintenance policy by more than 8%.
关键词：KeywordsPredictive maintenancedata-drivenclassificationsmart power applications
